Good morning.

Two weeks ago I named the cognitive layer as the part of your business agents read from, and the part almost nobody has built yet.

The response was the predictable one: people understood the idea and asked for the file structure.

This issue is the file structure. One folder. A small set of files. A seven-step protocol to install it this afternoon, plus the handoff that keeps it alive once your team is writing into it.

By the time you finish this issue you will have a working cognitive layer your agents can read from, and use to maximize whatever task you have them working on.

— Sam

Let’s get into it.

Most operators trying to fix the agent-context problem reach for a tool. A new Notion workspace. A vector database. A memory product. A Zapier flow that copies things between systems.

The problem with all of those is that agents do not work the way humans work. An agent reads files. It pulls them into a context window, processes them, writes output, and ends.

The fastest, cheapest, most portable way to give your agents shared context is a folder of plain markdown files that any model can read regardless of the platform you happen to be on this quarter.

A folder also survives platform cycles. If you migrate from Claude to whatever ships next year, the folder moves with you. The files do not care which model reads them. That portability is the whole point. You are building an asset, not a configuration.

One e-commerce operator I work with has a single shared folder his three agents all draw from. About 18 markdown files. Every model his team has tested has been able to read them. The folder is older than two of the agent platforms he has run his business on, and it will outlive the next two as well.

Here is the full tree to start from. You can grow it later, but resist the urge to start bigger than this.

A short tour before you start building.

CLAUDE.md sits at the root and holds the master context every agent loads. The name is becoming a cross-model standard, with AGENTS.md as the emerging alternative. Pick one of the two and stay consistent across the folder.

agent-team.md is the org chart of your agent layer. It describes who exists, what each one does, and what each one reads from.

The context/ folder holds business knowledge that is true regardless of which agent is running. ICPs, offers, voice, processes. Multiple agents read from these files, and you write them once.

The agents/ folder holds one subfolder per agent. Each subfolder has a SOUL.md for identity and a memory/ folder where the agent logs what it learned across runs.

The skills/ folder is the modular layer. Each skill is its own subfolder with a SKILL.md that any of your agents can load on demand. The Skill Fire issue covers how to build one, and that work is the right next step once the folder structure is in place.

The whole thing has to live somewhere your agents can read. A Claude project folder, a Cowork workspace, a GitHub repo synced to your agent platform, or a shared cloud folder, depending on where your agents run. Where it lives matters less than picking one location and not splitting across two.

Conventions that hold up under real use, not theory. Each file does one job.

The master context file. Every agent in your business should load it before doing anything else. Keep it under 1,000 words. If it grows past that, push detail down into context files and link to them.

What goes in:

CLAUDE.md is the index. The context files hold the substance.

Your agent org chart. One section per agent, with three lines each.

If you only have one agent today, this file is short. That is fine. The discipline of writing it down forces the question of how each new agent fits with the ones already running.

The deeper version of the one-paragraph in CLAUDE.md. History, key metrics, what is true about the business that any agent might need to reference.

Your ideal customer profiles. Who they are, what they want, what they hate, how they buy. If you have multiple ICPs, give each one a section. Agents drafting outreach, ads, or content all reference this file. Without it, every agent guesses.

Every offer you sell, with price, scope, deliverables, positioning, and common objections. Sales, marketing, and support agents all reference this file, so write it once and route everyone to it.

The voice guide. House style, banned language, sentence rules, and formatting preferences in one place. Include two or three examples of writing that hits the voice plus one example that misses, with notes on why. The miss is often more useful than the hit.

The repeatable workflows your business runs. How a deal gets qualified. How a piece of content gets shipped. How a refund gets handled. Agents reference these to know where they sit in the larger flow.

Per agent. Identity, role, behavior, hard limits.

A SOUL.md is what makes an agent feel like a member of the team rather than a chat session.

Where the agent (or you) logs decisions that came out of its work. Why a particular call was made. What the data said. What the agent flagged that turned out to matter.

This is the file that makes the agent compound. Without it, every run starts cold.

Where you log what happened after the decisions in decisions.md. Did the call work. Did the recommendation hold up. Where the agent was wrong.

The pairing of decisions and outcomes is what closes the loop. Over a few months, an agent reading both files starts to make calls that look like the ones that worked in production, not the ones that sounded reasonable in isolation.

Modular expertise. Each skill is a folder with a SKILL.md at the top. Multiple agents can load the same skill on demand. The Skill Fire issue is the deep dive on building these well.

Seven steps. One afternoon if you focus. The bar is a working folder your first agent can read from by end of day. Polish comes later.

Step 1. Pick the location. Choose where the folder will live. A Claude project. A Cowork workspace. A GitHub repo. A shared cloud folder synced to your agent platform. The right answer is wherever your agents already run. Create the empty skeleton: a cognitive-layer/ folder at the top, with the subfolders shown above. Empty files are fine.

Step 2. Write CLAUDE.md. Open the file at the root and write the master context. The business in one paragraph. The team. Current priorities. What is off-limits. Pointers to the context files you will fill in later. Twenty to thirty minutes if you keep it tight.

Step 3. Fill the context folder. Write business-overview.md, icps.md, offers.md, voice.md, and processes.md. Pull from your existing docs. Most operators have 60% of this written already, scattered across Google Drive and Notion. The work is consolidation, not creation. If a file feels thin, mark it [TODO: expand] and keep moving. The folder works at 70% completeness. Three weeks of polishing kills the project before it ships.

Step 4. Write agent-team.md. List every agent you currently have running, plus the next one or two you plan to build. Three lines each: what it does, what it reads from, what it produces. If you only have one agent today, write the file anyway. The structure forces the right question every time you add another.

Step 5. Build the first agent's folder. Pick the agent you use most. Create the subfolder under agents/. Write its SOUL.md. Create the empty memory/decisions.md and memory/outcomes.md files. The memory files start blank and get populated through the Friday ritual described below.

Step 6. Add one skill. Pick one recurring task that agent does. Build a skill for it: a folder under skills/, a SKILL.md with a trigger-style description and a structured body. The Skill Fire issue is the reference. One skill is enough for the first pass. Two more can come next week.

Step 7. Test the read path. Point your agent at the folder. Ask it a question that should require it to load something specific from the context files. Watch the run end-to-end. If the agent ignored a file you expected it to read, the issue is almost always in the SOUL.md (the agent does not know the file exists or where it lives) or in CLAUDE.md (the pointer is missing or unclear). Fix it, run again. The first time the folder works the way it should, the whole project clicks.

By the end of step seven you have a working cognitive layer. It will be 70% complete and that is the right place to start. The remaining 30% comes from the team writing into it.

If you would rather not run the seven-step protocol by hand, here is a Claude Skill that runs the whole conversation and writes the full folder at the end. Same output as the Protocol, with Claude doing the file creation while you answer the questions.

Three steps to install it:

The full SKILL.md to paste:

Run it once to install the folder. Run it again whenever you stand up a new agent and want it scaffolded the same way.

A folder built and abandoned is worse than a folder that never existed. Here is how the team angle works.

Assign one owner per file. CLAUDE.md is yours, the operator. Voice and processes might be owned by your head of operations. ICPs and offers might be owned by sales or marketing. Each owner is responsible for keeping their file current and accurate. Anyone on the team can suggest edits, but the owner ships the changes.

Give the team write access to the memory files. The decisions.md file is where your team logs why a call was made. Voice-to-text works. Rough prose works. The reader is an agent, not a copy editor. The point is to capture the reasoning before it disappears.

Pick one tool for editing. If it lives in Google Drive, edit in Google Drive. If it lives in GitHub, edit through your usual git flow. If it lives in a Cowork or Claude project, edit there. The fastest way to kill the folder is to have three teammates editing in three places.

Onboard new hires by walking them through the folder. Reading CLAUDE.md, agent-team.md, and the context files is faster than three weeks of asking the same questions in Slack. The folder doubles as your operating manual without you having to write a separate operating manual.

Your cognitive folder is the part of your business that ages well. 

Tools change and agents get rebuilt, and platforms come and go. 

The folder moves with you, and every week it becomes a little more useful than it was the week before.

If today's issue landed, Cortex is where you get more like this and where work gets done.

And Cortex is open for new subscribers:

Cortex is a monthly premium version of Bionic Business, helping entrepreneurs and business owners grow and scale with AI and Agents.

You only get access to the Cortex (Signals and Circuits) issues for the months you're subscribed to, not the archives.

You can cancel at any time. But there are no refunds, for any reason. All sales are final.

Get in now in time for the next batch of issues.

Until next time,
Sam Woods
The Editor

.